def __init__(self,
name,
pipeline_parser,
pipeline_config_key,
options_fn=None):
self.name = name
self._pipeline_parser = pipeline_parser
self._pipeline_config_key = pipeline_config_key.replace('-', '_')
self._parser = configargparse.get_argument_parser(
self.name,
prog='kiwi {}'.format(self.name),
add_help=False,
config_file_parser_class=configargparse.YAMLConfigFileParser,
ignore_unknown_config_file_keys=False,
)
self._parser.add(
'--config',
required=False,
is_config_file=True,
type=PathType(exists=True),
help='Load config file from path',
)
if options_fn is not None:
options_fn(self._parser)
def _add_predicting_data_file_opts(parser):
# Data options
group = parser.add_argument_group('data')
group.add_argument(
'--test-source',
type=PathType(exists=True),
required=True,
help='Path to validation source file',
)
group.add_argument(
'--test-target',
type=PathType(exists=True),
required=True,
help='Path to validation target file',
)
return group
def _add_data_options(data_parser):
group = data_parser.add_argument_group(
'PredEst data',
description='Predictor Estimator specific data ' 'options. (POSTECH)',
)
group.add(
'--extend-source-vocab',
type=PathType(exists=True),
help='Optionally load more data which is used only for vocabulary '
'creation. Path to additional Data'
'(Predictor)',
)
group.add(
'--extend-target-vocab',
type=PathType(exists=True),
help='Optionally load more data which is used only for vocabulary '
'creation. Path to additional Data'
'(Predictor)',
)
def add_predicting_data_file_opts(parser):
# Data options
group = parser.add_argument_group('data')
group.add_argument(
'--test-source',
type=PathType(exists=True),
required=True,
help='Path to validation source file',
)
group.add_argument(
'--test-target',
type=PathType(exists=True),
required=True,
help='Path to validation target file',
)
group.add_argument(
'--test-alignments',
type=PathType(exists=True),
help='Path to test alignments between source and target.',
)
group.add_argument(
'--test-source-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for source',
)
group.add_argument(
'--test-target-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for target',
)
group.add_argument(
'--test-target-parse',
type=PathType(exists=True),
help='Path to test dependency parsing file for target (tabular format)',
)
group.add_argument(
'--test-target-ngram',
type=PathType(exists=True),
help='Path to test highest order ngram file for target (tabular '
'format)',
) # noqa
group.add_argument(
'--test-target-stacked',
type=PathType(exists=True),
help='Path to test stacked predictions file for target (tabular '
'format)',
) # noqa
return group
def jackknife_opts(parser):
# Training loop options
group = parser.add_argument_group('jackknifing')
group.add(
'--splits',
required=False,
type=int,
default=5,
help='Jackknife with X folds.',
)
group.add(
'--train-config',
required=False,
type=PathType(exists=True),
help='Path to config file with model parameters.',
)
def add_training_options(training_parser):
add_pretraining_options(training_parser)
group = training_parser.add_argument_group(
'predictor-estimator training',
description='Predictor Estimator (POSTECH). These settings are used '
' to train the Predictor. They will be ignored if training a '
' Predictor-Estimator and the `load-model` flag is set.',
)
group.add_argument(
'--start-stop',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Append start and stop symbols to estimator feature sequence.',
)
group.add_argument(
'--predict-gaps',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Predict Gap Tags. Requires `train-gap-tags`, `valid-'
'gap-tags` to be set.',
)
group.add_argument(
'--predict-target',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=True,
help='Predict Target Tags. Requires `train-target-tags`, `valid-'
'target-tags` to be set.',
)
group.add_argument(
'--predict-source',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Predict Source Tags. Requires `train-source-tags`, `valid-'
'source-tags` to be set.',
)
group.add_argument(
'--load-pred-source',
type=PathType(exists=True),
help='If set, model architecture and vocabulary parameters are '
'ignored. Load pretrained predictor tgt->src.',
)
group.add_argument(
'--load-pred-target',
type=PathType(exists=True),
help='If set, model architecture and vocabulary parameters are '
'ignored. Load pretrained predictor src->tgt.',
)
group.add_argument(
'--rnn-layers-est', type=int, default=2, help='Layers in Estimator RNN'
)
group.add_argument(
'--dropout-est', type=float, default=0.0, help='Dropout in estimator'
)
group.add_argument(
'--hidden-est',
type=int,
default=100,
help='Size of hidden layers in LSTM',
)
group.add_argument(
'--mlp-est',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help="""Pass the Estimator input through a linear layer
reducing dimensionality before RNN.""",
)
group.add_argument(
'--sentence-level',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help="""Predict Sentence Level Scores.
Requires setting `train-sentence-scores, valid-sentence-scores`""",
)
group.add_argument(
'--sentence-ll',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help="""Use probabilistic Loss for sentence scores instead of
squared error. If set, the model will output mean and variance of
a truncated Gaussian distribution over the interval [0, 1], and use
the NLL of ground truth `hter` as the loss.
This seems to improve performance, and gives you uncertainty estimates
for sentence level predictions as a byproduct.
If `sentence-level == False`, this is without effect.
""",
)
group.add_argument(
'--sentence-ll-predict-mean',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help="""If `sentence-ll == True`, by default the prediction for `hter`
will be the mean of the Guassian /before truncation/. After truncation,
this will be the mode of the distribution, but not the mean as
truncated Gaussian is skewed to one side. set this to `True` to use
the True mean after truncation for prediction.
""",
)
group.add_argument(
'--use-probs',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Predict scores as product/sum of word level probs',
)
group.add_argument(
'--binary-level',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help="""Predict binary sentence labels indicating `hter == 0.0`
Requires setting `train-sentence-scores`, `valid-sentence-scores`""",
)
group.add_argument(
'--token-level',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help="""Continue training the predictor on the postedited text.
If set, will do an additional forward pass through the predictor
Using the SRC, PE pair and add the `Predictor` loss for the tokens
in the postedited text PE. Recommended if you have access to PE.
Requires setting `train-pe`, `valid-pe`""",
)
group.add_argument(
'--target-bad-weight',
type=float,
default=3.0,
help='Relative weight for target bad labels.',
)
group.add_argument(
'--gaps-bad-weight',
type=float,
default=3.0,
help='Relative weight for gaps bad labels.',
)
group.add_argument(
'--source-bad-weight',
type=float,
default=3.0,
help='Relative weight for source bad labels.',
)
def _add_training_data_file_opts(parser):
# Data options
group = parser.add_argument_group('data')
group.add_argument(
'--train-source',
type=PathType(exists=True),
required=True,
help='Path to training source file',
)
group.add_argument(
'--train-target',
type=PathType(exists=True),
# required=True,
help='Path to training target file',
)
group.add_argument(
'--train-source-tags',
type=PathType(exists=True),
help='Path to validation label file for source (WMT18 format)',
)
group.add_argument(
'--train-target-tags',
type=PathType(exists=True),
help='Path to validation label file for target',
)
group.add_argument(
'--train-pe',
type=PathType(exists=True),
help='Path to file containing post-edited target.',
)
group.add_argument(
'--train-sentence-scores',
type=PathType(exists=True),
help='Path to file containing sentence level scores.',
)
valid_group = parser.add_argument_group('validation data')
valid_group.add_argument(
'--split',
type=float,
help='Split Train dataset in case that no validation set is given.',
)
valid_group.add_argument(
'--valid-source',
type=PathType(exists=True),
# required=True,
help='Path to validation source file',
)
valid_group.add_argument(
'--valid-target',
type=PathType(exists=True),
# required=True,
help='Path to validation target file',
)
valid_group.add_argument(
'--valid-alignments',
type=str,
# required=True,
help='Path to valid alignments between source and target.',
)
valid_group.add_argument(
'--valid-source-tags',
type=PathType(exists=True),
help='Path to validation label file for source (WMT18 format)',
)
valid_group.add_argument(
'--valid-target-tags',
type=PathType(exists=True),
help='Path to validation label file for target',
)
valid_group.add_argument(
'--valid-pe',
type=PathType(exists=True),
help='Path to file containing postedited target.',
)
valid_group.add_argument(
'--valid-sentence-scores',
type=PathType(exists=True),
help='Path to file containing sentence level scores.',
)
def _add_vocabulary_opts(parser):
group = parser.add_argument_group(
'vocabulary options',
description='Options for loading vocabulary from a previous run. '
'This is used for e.g. training a source predictor via predict-'
'inverse: True ; If set, other vocab options are ignored',
)
group.add_argument(
'--source-vocab-size',
type=int,
default=None,
help='Size of the source vocabulary.',
)
group.add_argument(
'--target-vocab-size',
type=int,
default=None,
help='Size of the target vocabulary.',
)
group.add_argument(
'--source-vocab-min-frequency',
type=int,
default=1,
help='Min word frequency for source vocabulary.',
)
group.add_argument(
'--target-vocab-min-frequency',
type=int,
default=1,
help='Min word frequency for target vocabulary.',
)
"""start:给predictor-estimator添加source-embeddings和target-embeddings"""
group.add_argument(
'--embeddings-format',
type=str,
default='polyglot',
choices=['polyglot', 'word2vec', 'fasttext', 'glove', 'text'],
help='Word embeddings format. '
'See README for specific formatting instructions.',
)
group.add_argument(
'--embeddings-binary',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Load embeddings stored in binary.',
)
group.add_argument(
'--source-embeddings',
type=PathType(exists=True),
help='Path to word embeddings file for source.',
)
group.add_argument(
'--target-embeddings',
type=PathType(exists=True),
help='Path to word embeddings file for target.',
)
"""end"""
def add_training_data_file_opts(parser):
# Data options
group = parser.add_argument_group('data')
group.add_argument(
'--train-source',
type=PathType(exists=True),
required=True,
help='Path to training source file',
)
group.add_argument(
'--train-target',
type=PathType(exists=True),
required=True,
help='Path to training target file',
)
group.add_argument(
'--train-alignments',
type=str,
required=True,
help='Path to train alignments between source and target.',
)
group.add_argument(
'--train-source-tags',
type=PathType(exists=True),
help='Path to training label file for source (WMT18 format)',
)
group.add_argument(
'--train-target-tags',
type=PathType(exists=True),
help='Path to training label file for target',
)
group.add_argument(
'--valid-source',
type=PathType(exists=True),
required=True,
help='Path to validation source file',
)
group.add_argument(
'--valid-target',
type=PathType(exists=True),
required=True,
help='Path to validation target file',
)
group.add_argument(
'--valid-alignments',
type=str,
required=True,
help='Path to valid alignments between source and target.',
)
group.add_argument(
'--valid-source-tags',
type=PathType(exists=True),
help='Path to validation label file for source (WMT18 format)',
)
group.add_argument(
'--valid-target-tags',
type=PathType(exists=True),
help='Path to validation label file for target',
)
group.add_argument(
'--valid-source-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for source',
)
group.add_argument(
'--valid-target-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for target',
)
return group
def add_vocabulary_opts(parser):
group = parser.add_argument_group('vocabulary options')
group.add_argument(
'--source-vocab-size',
type=int,
default=None,
help='Size of the source vocabulary.',
)
group.add_argument(
'--target-vocab-size',
type=int,
default=None,
help='Size of the target vocabulary.',
)
group.add_argument(
'--source-vocab-min-frequency',
type=int,
default=1,
help='Min word frequency for source vocabulary.',
)
group.add_argument(
'--target-vocab-min-frequency',
type=int,
default=1,
help='Min word frequency for target vocabulary.',
)
group.add_argument(
'--keep-rare-words-with-embeddings',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Keep words that occur less then min-frequency but '
'are in embeddings vocabulary.',
)
group.add_argument(
'--add-embeddings-vocab',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Add words from embeddings vocabulary to source/target '
'vocabulary.',
)
group.add_argument(
'--embeddings-format',
type=str,
default='polyglot',
choices=['polyglot', 'word2vec', 'fasttext', 'glove', 'text'],
help='Word embeddings format. '
'See README for specific formatting instructions.',
)
group.add_argument(
'--embeddings-binary',
type=lambda x: bool(strtobool(x)),
nargs='?',
const=True,
default=False,
help='Load embeddings stored in binary.',
)
group.add_argument(
'--source-embeddings',
type=PathType(exists=True),
help='Path to word embeddings file for source.',
)
group.add_argument(
'--target-embeddings',
type=PathType(exists=True),
help='Path to word embeddings file for target.',
)
return group
parser = configargparse.get_argument_parser('search')
parser.add_argument(
'-e',
'--experiment-name',
required=False,
help='MLflow will log this run under this experiment name, '
'which appears as a separate section in the UI. It '
'will also be used in some messages and files.',
)
parser.add(
'-c',
'--config',
required=True,
is_config_file=False,
type=PathType(exists=True),
help='Load config file from path',
)
group = parser.add_argument_group('models')
group.add_argument('model_name', choices=Model.subclasses.keys())
def get_action(option):
for action in train.parser._actions:
if option in train.parser.get_possible_config_keys(action):
return action
return None
def split_options(options):
def add_training_data_file_opts(parser):
# Data options
group = parser.add_argument_group('data')
group.add_argument(
'--train-source',
type=PathType(exists=True),
help='Path to training source file',
)
group.add_argument(
'--train-target',
type=PathType(exists=True),
help='Path to training target file',
)
group.add_argument(
'--train-alignments',
type=str,
help='Path to train alignments between source and target.',
)
group.add_argument(
'--train-source-tags',
type=PathType(exists=True),
help='Path to validation label file for source (WMT18 format)',
)
group.add_argument(
'--train-target-tags',
type=PathType(exists=True),
help='Path to validation label file for target',
)
group.add_argument(
'--train-source-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for source',
)
group.add_argument(
'--train-target-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for target',
)
group.add_argument(
'--train-target-parse',
type=PathType(exists=True),
help='Path to training dependency parsing file for target (tabular '
'format)',
)
group.add_argument(
'--train-target-ngram',
type=PathType(exists=True),
help='Path to training highest order ngram file for target (tabular '
'format)',
)
group.add_argument(
'--train-target-stacked',
type=PathType(exists=True),
help='Path to training stacked predictions file for target (tabular '
'format)',
)
group = parser.add_argument_group('validation data')
group.add_argument(
'--valid-source',
type=PathType(exists=True),
# required=True,
help='Path to validation source file',
)
group.add_argument(
'--valid-target',
type=PathType(exists=True),
# required=True,
help='Path to validation target file',
)
group.add_argument(
'--valid-alignments',
type=str,
# required=True,
help='Path to valid alignments between source and target.',
)
group.add_argument(
'--valid-source-tags',
type=PathType(exists=True),
help='Path to validation label file for source (WMT18 format)',
)
group.add_argument(
'--valid-target-tags',
type=PathType(exists=True),
help='Path to validation label file for target',
)
group.add_argument(
'--valid-source-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for source',
)
group.add_argument(
'--valid-target-pos',
type=PathType(exists=True),
help='Path to training PoS tags file for target',
)
group.add_argument(
'--valid-target-parse',
type=PathType(exists=True),
help='Path to validation dependency parsing file for target (tabular '
'format)',
)
group.add_argument(
'--valid-target-ngram',
type=PathType(exists=True),
help='Path to validation highest order ngram file for target (tabular '
'format)',
)
group.add_argument(
'--valid-target-stacked',
type=PathType(exists=True),
help='Path to validation stacked predictions file for target (tabular '
'format)',
)
def evaluate_opts(parser):
# Evaluation options
group = parser.add_argument_group("Evaluation of WMT Quality Estimation")
group.add_argument(
"--type",
help="Input type for prediction file",
choices=["probs", "tags"],
type=str,
default="probs",
)
group.add_argument(
"--format",
help="Input format for gold files",
choices=["wmt17", "wmt18"],
type=str,
default="wmt17",
)
group.add_argument(
"--pred-format",
help="Input format for predicted files. Defaults to the same as "
"--format.",
choices=["wmt17", "wmt18"],
type=str,
default="wmt18",
)
group.add_argument(
"--sents-avg",
help="Obtain scores for sentences by averaging over tags or "
"probabilities.",
choices=["probs", "tags"],
type=str,
# default=None
)
# Gold files.
group.add_argument(
"--gold-sents",
help="Sentences gold standard. ",
type=PathType(exists=True),
required=False,
)
group.add_argument(
"--gold-target",
help="Target tags gold standard, or target and gaps "
'if format == "wmt18".',
type=PathType(exists=True),
required=False,
)
group.add_argument(
"--gold-source",
help="Source tags gold standard.",
type=PathType(exists=True),
required=False,
)
group.add_argument(
"--gold-cal",
help="Target Tags to calibrate.",
type=PathType(exists=True),
required=False,
)
# Prediction Files
group.add_argument(
"--input-dir",
help="Directory with prediction files generated by predict pipeline. "
"Setting this argument will evaluate all predictions for "
"which a gold file is set.",
nargs="+",
type=PathType(exists=True),
# required=True
)
group.add_argument(
"--pred-sents",
help="Sentences HTER predictions.",
type=PathType(exists=True),
nargs="+",
required=False,
)
group.add_argument(
"--pred-target",
help="Target predictions; can be tags or probabilities (of BAD). "
"See --type.",
type=PathType(exists=True),
nargs="+",
required=False,
)
group.add_argument(
"--pred-gaps",
help="Gap predictions; can be tags or probabilities (of BAD). "
"(see --type). Use this option for files that only contain gap "
"tags.",
type=PathType(exists=True),
nargs="+",
required=False,
)
group.add_argument(
"--pred-source",
help="Source predictions. can be tags or probabilities (of BAD). "
" See --type.",
type=PathType(exists=True),
nargs="+",
required=False,
)
group.add_argument(
"--pred-cal",
help="Target Predictions to calibrate.",
type=PathType(exists=True),
required=False,
)